Comparison of regression techniques to predict attractiveness from facial colour cues

Poster Presentation 53.321: Tuesday, May 23, 2023, 8:30 am – 12:30 pm, Banyan Breezeway
Session: Face Perception: Models

There is a Poster PDF for this presentation, but you must be a current member or registered to attend VSS 2023 to view it.
Please go to your Account Home page to register.

Yan Lu1 (), Kaida Xiao1,2, Jie Yang3, Michael Pointer1, Changjun Li2, Sophie Wuerger4; 1Leeds Institute of Textile and Colour, University of Leeds, 2University of Science and Technology Liaoning, 3Beijing Institute of Graphic Communication, 4University of Liverpool

Various facial colour cues (average/local skin colour, colour contrasts, colour variations, etc.) were identified as valid predictors of facial attractiveness. Conventional studies on single colour variables simplified the complex nature of attractiveness judgement on real human faces. However, predicting attractiveness from various colour cues is difficult due to the high number of candidate variables and their correlations. In this study, multivariate statistical techniques and machine learning (ML) algorithms were utilized to model the relationship between facial attractiveness and a large number of colour variables using Chinese samples. One hundred images of real human faces were used as the experimental materials, with the colour rigorously controlled to represent the naturally occurring facial colour variations in Chinese populations. Two separate attractiveness evaluation data were collected through psychophysical experiments as training and testing dataset, respectively. We proposed eight strategies for robust regression of the high-dimensional dataset based on three techniques: subset selection (forward, backward stepwise), dimension reduction (principal component regression, partial least-squares regression), and regularization (Ridge, Lasso, Elastic Net regression). Model performance was evaluated by the predictive accuracy, the goodness of fit, and the selection of colour predictors. Results showed the out-of-sample root-mean-square error for dimension reduction and regularization methods was better than the classical least-squares. The best ML algorithm predicted facial attractiveness within 0.67 points on a 7-point scale. Different predictors were selected depending on methods but several common predictors were revealed as important features including skin lightness, overall colour variation, and colour contrast around eyebrows. Here we evaluated statistical and ML algorithms for utilizing facial colour cues for attractiveness prediction based on realistic skin models. From the perspective of both well-predicting and interpretable, ML techniques with feature selection were recommended for attractiveness modelling. Our results also demonstrated the importance of colour to facial attractiveness which is comparable to those structural features.